TEDDY AI Model Predicts Pediatric Disease Risk from EHR Data

Pediatric electronic health records (EHRs) contain rich, developmentally structured clinical trajectories, yet their potential for generative healthcare foundation models has been largely untapped. This paper introduces TEDDY (Temporal Event Decoder for Disease in Youth), a compact 1.84-million-parameter decoder transformer. TEDDY was trained on a massive dataset of approximately 73 million ICD-10 diagnoses from 1.6 million children at a single pediatric institution, learning to model longitudinal diagnosis trajectories and visit timing. TEDDY's predictions for disease onset, made before visit codes were revealed, were evaluated against sex- and age-matched controls across 797 disease-onset tasks spanning 16 ICD-10 chapters. It achieved a median AUC of 72.0%, significantly outperforming traditional machine learning baselines (DenseNet, CNN, RNN, LSTM) on 96-99% of tasks. Its performance was particularly strong for lower-prevalence diagnoses, with 90% of the rarest conditions showing confidence intervals above chance. The model demonstrated predictive signal more than two years before the first recorded diagnosis, with robust median AUCs. For specific benchmarks like asthma and ADHD, TEDDY achieved AUCs of 79.3% and 84.7%, surpassing even a general-purpose language model three orders of magnitude larger. While visit-timing predictions showed a 3.0-day mean absolute error, the overall results establish pediatric diagnostic histories as a viable substrate for compact generative models capable of broad, rare-disease, and long-horizon risk forecasting without requiring population-scale data or massive models.